2.1 Accessing the data in R

The pRolocdata package is a R Bioconductor experiment package that collects mass spectrometry-based spatial/organelle and protein-complex datasets from published experiments. Each data is stored in a container called a MSnSet instance (see the MSnbase for details) which allows users to work seamlessly with the R pRoloc and pRolocGUI software for spatial proteomics data analysis and visualisation.

The first step is to install the pRoloc package from Bioconductor and the devtools package to access pRolocdata on Github

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install("pRoloc")
BiocManager::install("devtools")

## we also use the knitr and VennDiagram, ggplot2 packages in this 
## vignette so please  install if you do not have them 
BiocManager::install(c("knitr", "VennDiagram", "ggplot2", "dplyr", "kableExtra", 
                       "coda", "reshape2", "circlize", "ggalluvial"))

Note: We can install the pRolocdata package from Bioconductor using the same method e.g. BiocManager::install("pRolocdata") however we require version >=1.29.1 which is not yet available on Bioconductor until the next stable release (Thursday 20th May). In the meantime we can access it via Github using the R package devtools using the below commands.

library("devtools")
install_github("https://github.com/lgatto/pRolocdata")

Now we can load the packages by calling library function in R

library("pRoloc")
library("pRolocdata")

and subsequently we can now access all the datasets in pRolocdata. To list all available datasets we use the data function.

data(package = "pRolocdata")

We see there are 27 datasets (including the PSM and protein level data) associated with the manuscript [1], 20 from the hyperLOPIT data anlysis and 7 from the LPS timecourse [1]. For more information on these datasets, we can type

?thpLOPIT_lps_mulvey2021
?lpsTimecourse_mulvey2021

Figure 1. Screenshot of the documentation files in R for the THP LOPIT and LPS timecourse datasets

Each dataset can be loaded using the data function. For example, to load the final protein level LOPIT datasets for the unstimulated and 12 hour LPS stimulated datasets,

data("thpLOPIT_unstimulated_mulvey2021")
thpLOPIT_unstimulated_mulvey2021

data("thpLOPIT_lps_mulvey2021")
thpLOPIT_lps_mulvey2021

In the next section we introduce the data, experimental design and brief overview of the experimental protocol before walking users through the downstream analysis.

3.1 The hyperLOPIT workflow

Using the hyperLOPIT proteomics platform [4, 5] we obtained spatial maps that capture the steady-state distribution of thousands of proteins in the THP-1 human leukaemia cell line. Experiments were conducted in triplicate and the samples were analysed and collected when cells were (1) unstimulated and then (2) at 12 hours following stimulation with LPS.

The hyperLOPIT method combines biochemical cell fractionation with multiplexed high-resolution mass-spectrometry [4, 5], the full protocol and experimental design can be found in [1]. Briefly, TMT labelling was conducted as described in [5] and 20 fractions including the cytosol-enriched fraction were labelled per gradient, by combining two TMT 10-plex experiments in an interleaved labelling design to capture as much subcellular diversity as possible. The experimental design for each dataset is stored in the pData slot of each dataset.

To access the full experimental design use the pData function

## TMT experiment data for the unstimulated hyperLOPIT data
pData(thpLOPIT_unstimulated_mulvey2021)

##                               Tag    Treatment Replicate Set Fraction
## unstim_rep1_set1_126_cyto     126 Unstimulated         1   1     cyto
## unstim_rep2_set1_126_cyto     126 Unstimulated         2   1     cyto
## unstim_rep3_set1_126_cyto     126 Unstimulated         3   1     cyto
## unstim_rep1_set1_127N_F1.4   127N Unstimulated         1   1     F1.4
## unstim_rep2_set1_127N_F1.6   127N Unstimulated         2   1     F1.6
## unstim_rep3_set1_127N_F1.6   127N Unstimulated         3   1     F1.6
## unstim_rep1_set2_126_F5.6     126 Unstimulated         1   2     F5.6
## unstim_rep1_set1_127C_F7     127C Unstimulated         1   1       F7
## unstim_rep2_set2_127N_F7.8   127N Unstimulated         2   2     F7.8
## unstim_rep3_set2_126_F7.8     126 Unstimulated         3   2     F7.8
## unstim_rep1_set2_127N_F8     127N Unstimulated         1   2       F8
## unstim_rep1_set1_128N_F9     128N Unstimulated         1   1       F9
## unstim_rep3_set1_127C_F9     127C Unstimulated         3   1       F9
## unstim_rep2_set1_127C_F9.10  127C Unstimulated         2   1    F9.10
## unstim_rep1_set2_127C_F10    127C Unstimulated         1   2      F10
## unstim_rep3_set2_127N_F10    127N Unstimulated         3   2      F10
## unstim_rep1_set1_128C_F11    128C Unstimulated         1   1      F11
## unstim_rep3_set1_128N_F11    128N Unstimulated         3   1      F11
## unstim_rep2_set2_127C_F11.12 127C Unstimulated         2   2   F11.12
## unstim_rep1_set2_128N_F12    128N Unstimulated         1   2      F12
## unstim_rep2_set1_128N_F12    128N Unstimulated         2   1      F12
## unstim_rep3_set2_127C_F12    127C Unstimulated         3   2      F12
## unstim_rep1_set1_129N_F13    129N Unstimulated         1   1      F13
## unsti_rep2_set2_128N_F13     128N Unstimulated         2   2      F13
## unstim_rep3_set1_128C_F13    128C Unstimulated         3   1      F13
## unstim_rep1_set2_128C_F14    128C Unstimulated         1   2      F14
## unstim_rep2_set1_128C_F14    128C Unstimulated         2   1      F14
## unstim_rep3_set2_128N_F14    128N Unstimulated         3   2      F14
## unstim_rep1_set1_129C_F15    129C Unstimulated         1   1      F15
## unstim_rep2_set2_128C_F15    128C Unstimulated         2   2      F15
## unstim_rep3_set1_129N_F15    129N Unstimulated         3   1      F15
## unstim_rep1_set2_129N_F16    129N Unstimulated         1   2      F16
## unstim_rep2_set1_129N_F16    129N Unstimulated         2   1      F16
## unstim_rep3_set2_128C_F16    128C Unstimulated         3   2      F16
## unstim_rep1_set1_130N_F17    130N Unstimulated         1   1      F17
## unstim_rep2_set2_129N_F17    129N Unstimulated         2   2      F17
## unstim_rep3_set1_129C_F17    129C Unstimulated         3   1      F17
## unstim_rep1_set2_129C_F18    129C Unstimulated         1   2      F18
## unstim_rep2_set1_129C_F18    129C Unstimulated         2   1      F18
## unstim_rep3_set2_129N_F18    129N Unstimulated         3   2      F18
## unstim_rep1_set1_130C_F19    130C Unstimulated         1   1      F19
## unstim_rep2_set2_129C_F19    129C Unstimulated         2   2      F19
## unstim_rep3_set1_130N_F19    130N Unstimulated         3   1      F19
## unstim_rep1_set2_130N_F20    130N Unstimulated         1   2      F20
## unstim_rep2_set1_130N_F20    130N Unstimulated         2   1      F20
## unstim_rep3_set2_129C_F20    129C Unstimulated         3   2      F20
## unstim_rep1_set1_131_F21      131 Unstimulated         1   1      F21
## unstim_rep2_set2_130N_F21    130N Unstimulated         2   2      F21
## unstim_rep3_set1_130C_F21    130C Unstimulated         3   1      F21
## unstim_rep1_set2_130C_F22    130C Unstimulated         1   2      F22
## unstim_rep2_set1_130C_F22    130C Unstimulated         2   1      F22
## unstim_rep3_set2_130N_F22    130N Unstimulated         3   2      F22
## unstim_rep2_set2_130C_F23    130C Unstimulated         2   2      F23
## unstim_rep3_set1_131_F23      131 Unstimulated         3   1      F23
## unstim_rep1_set2_131_F24      131 Unstimulated         1   2      F24
## unstim_rep2_set1_131_F24      131 Unstimulated         2   1      F24
## unstim_rep3_set2_130C_F24    130C Unstimulated         3   2      F24
## unstim_rep2_set2_131_F25      131 Unstimulated         2   2      F25
## unstim_rep3_set2_131_F26      131 Unstimulated         3   2      F26

## TMT labelling scheme for the LPS hyperLOPIT data
pData(thpLOPIT_lps_mulvey2021)

##                           Tag Treatment Replicate Set Fraction
## lps_rep1_set1_126_cyto    126       LPS         1   1     cyto
## lps_rep2_set1_126_cyto    126       LPS         2   1     cyto
## lps_rep3_set1_126_cyto    126       LPS         3   1     cyto
## lps_rep1_set1_127N_F1.4  127N       LPS         1   1     F1.4
## lps_rep2_set1_127N_F1.6  127N       LPS         2   1     F1.6
## lps_rep3_set1_127N_F1.6  127N       LPS         3   1     F1.6
## lps_rep1_set2_126_F5.7    126       LPS         1   2     F5.7
## lps_rep2_set2_126_F7.8    126       LPS         2   2     F7.8
## lps_rep3_set2_126_F7.8    126       LPS         3   2     F7.8
## lps_rep1_set1_127C_F8    127C       LPS         1   1       F8
## lps_rep1_set2_127N_F9    127N       LPS         1   2       F9
## lps_rep3_set1_127C_F9    127C       LPS         3   1       F9
## lps_rep2_set1_127C_F9.10 127C       LPS         2   1    F9.10
## lps_rep1_set1_128N_F10   128N       LPS         1   1      F10
## lps_rep3_set2_127N_F10   127N       LPS         3   2      F10
## lps_rep1_set2_127C_F11   127C       LPS         1   2      F11
## lps_rep2_set2_127N_F11   127N       LPS         2   2      F11
## lps_rep3_set1_128N_F11   128N       LPS         3   1      F11
## lps_rep1_set1_128C_F12   128C       LPS         1   1      F12
## lps_rep2_set1_128N_F12   128N       LPS         2   1      F12
## lps_rep3_set2_127C_F12   127C       LPS         3   2      F12
## lps_rep1_set2_128N_F13   128N       LPS         1   2      F13
## lps_rep2_set2_127C_F13   127C       LPS         2   2      F13
## lps_rep3_set1_128C_F13   128C       LPS         3   1      F13
## lps_rep1_set1_129N_F14   129N       LPS         1   1      F14
## lps_rep2_set1_128C_F14   128C       LPS         2   1      F14
## lps_rep3_set2_128N_F14   128N       LPS         3   2      F14
## lps_rep1_set2_128C_F15   128C       LPS         1   2      F15
## lps_rep2_set2_128N_F15   128N       LPS         2   2      F15
## lps_rep3_set1_129N_F15   129N       LPS         3   1      F15
## lps_rep1_set1_129C_F16   129C       LPS         1   1      F16
## lps_rep2_set1_129N_F16   129N       LPS         2   1      F16
## lps_rep3_set2_128C_F16   128C       LPS         3   2      F16
## lps_rep1_set2_129N_F17   129N       LPS         1   2      F17
## lps_rep2_set2_128C_F17   128C       LPS         2   2      F17
## lps_rep3_set1_129C_F17   129C       LPS         3   1      F17
## lps_rep1_set1_130N_F18   130N       LPS         1   1      F18
## lps_rep2_set1_129C_F18   129C       LPS         2   1      F18
## lps_rep3_set2_129N_F18   129N       LPS         3   2      F18
## lps_rep1_set2_129C_F19   129C       LPS         1   2      F19
## lps_rep2_set2_129N_F19   129N       LPS         2   2      F19
## lps_rep3_set1_130N_F19   130N       LPS         3   1      F19
## lps_rep1_set1_130C_F20   130C       LPS         1   1      F20
## lps_rep2_set1_130N_F20   130N       LPS         2   1      F20
## lps_rep3_set2_129C_F20   129C       LPS         3   2      F20
## lps_rep1_set2_130N_F21   130N       LPS         1   2      F21
## lps_rep2_set2_129C_F21   129C       LPS         2   2      F21
## lps_rep3_set1_130C_F21   130C       LPS         3   1      F21
## lps_rep1_set1_131_F22     131       LPS         1   1      F22
## lps_rep2_set1_130C_F22   130C       LPS         2   1      F22
## lps_rep3_set2_130N_F22   130N       LPS         3   2      F22
## lps_rep1_set2_130C_F23   130C       LPS         1   2      F23
## lps_rep2_set2_130N_F23   130N       LPS         2   2      F23
## lps_rep3_set1_131_F23     131       LPS         3   1      F23
## lps_rep2_set1_131_F24     131       LPS         2   1      F24
## lps_rep3_set2_130C_F24   130C       LPS         3   2      F24
## lps_rep1_set2_131_F25     131       LPS         1   2      F25
## lps_rep2_set2_130C_F25   130C       LPS         2   2      F25
## lps_rep3_set2_131_F26     131       LPS         3   2      F26

As described in [1] and [5] the data was processed with Proteome Discoverer v2.1 (Thermo Fisher Scientific) and Mascot server v2.3.02 (Matrix Science) using the SwissProt sequence database for Homo sapiens (www.uniprot.org in November 2016) and the cRAP database (common Repository for Adventitious Proteins, https://www.thegpm.org/crap). Standard filtering of the raw data was conducted as per [1] and [5] to remove contaminants and low abundant PSMs. A maximum of two missing values per TMT experiment was allowed and the data was directly exported from Proteome Discoverer at the PSM level for downstream analysis in R.

The experiments were done in triplicate for each condition and each replicate contains 2 x TMT 10-plex sets (which we denote as “set 1” and “set 2”). Thus, in total we conducted a total of 12 experiments, 6 per condition wherein we had 2 sets per replicate containing a total of 20 fractions/channels per replicate. One TMT channel was removed (TMT tag 126 in replicate 2, set 2 in the unstimulated) due to erroneous labelling of insoluble material during the sample preparation for this specific tag. The “Average Reporter S/N” value was recalculated for the nine remaining channels in the corresponding 10plex set and PSMs with a value less than 9.0 were discarded (please see Data Processing in the Supplementary Information of [1]).

3.2 PSM level data processing

## Unstimulated data
data("psms_thpLOPIT_unstim_rep1_set1")
data("psms_thpLOPIT_unstim_rep1_set2")

data("psms_thpLOPIT_unstim_rep2_set1")
data("psms_thpLOPIT_unstim_rep2_set2")

data("psms_thpLOPIT_unstim_rep3_set1")
data("psms_thpLOPIT_unstim_rep3_set2")

## 12h post LPS-stimulated
data("psms_thpLOPIT_lps_rep1_set1")
data("psms_thpLOPIT_lps_rep1_set2")

data("psms_thpLOPIT_lps_rep2_set1")
data("psms_thpLOPIT_lps_rep2_set2")

data("psms_thpLOPIT_lps_rep3_set1")
data("psms_thpLOPIT_lps_rep3_set2")

psms <- MSnSetList(
  list(
    psms_thpLOPIT_unstim_rep1_set1,
    psms_thpLOPIT_unstim_rep1_set2,
    psms_thpLOPIT_unstim_rep2_set1,
    psms_thpLOPIT_unstim_rep2_set2,
    psms_thpLOPIT_unstim_rep3_set1,
    psms_thpLOPIT_unstim_rep3_set2,
    psms_thpLOPIT_lps_rep1_set1,
    psms_thpLOPIT_lps_rep1_set2,
    psms_thpLOPIT_lps_rep2_set1,
    psms_thpLOPIT_lps_rep2_set2,
    psms_thpLOPIT_lps_rep3_set1,
    psms_thpLOPIT_lps_rep3_set2
  )
)

## add list names for each dataset
.nams <- paste0("Rep", rep(1:3, each = 2), "_set", rep(1:2, 3))
names(psms) <- c(paste0(.nams, "_unstim"), paste0(.nams, "_lps"))

sapply(psms, nrow)

## Rep1_set1_unstim Rep1_set2_unstim Rep2_set1_unstim Rep2_set2_unstim 
##            44137            41860            39297            37919 
## Rep3_set1_unstim Rep3_set2_unstim    Rep1_set1_lps    Rep1_set2_lps 
##            48199            44802            40994            63844 
##    Rep2_set1_lps    Rep2_set2_lps    Rep3_set1_lps    Rep3_set2_lps 
##            43799            44055            46805            45600

for (i in seq(psms)) {
  par(las = 2, oma = c(10, 1, 1, 1))
  naplot(psms[[i]], col = "black", las = 2, reorderColumns = FALSE, 
         main = names(psms)[i], cex.axis = 2)
}

psms.imputed <- lapply(psms, function(z) impute(z, method = "MinDet"))

psms.imputed.norm <- lapply(psms.imputed, function(z) normalise(z,  "sum"))

prots <- lapply(psms.imputed.norm, function(z) 
  combineFeatures(z, method = "median",
                  groupBy = fData(z)[, "Master.Protein.Accessions"]))

## become combining all protein sets we tidy up the feature labels of the dataset
ll <- paste0(rep(c("unst", "lps"), 1, each = 6), ".r",
             rep(1:3, 1, each = 2), ".s", 1:2)
for (i in seq(prots)) {
  prots@x[[i]] <- updateFvarLabels(prots[[i]], label = ll[i], sep = "_")
}

## combine TMT sets and examine each replicate for each condition
unst_r1 <- filterNA(MSnbase::combine(prots[[1]], prots[[2]]))
unst_r2 <- filterNA(MSnbase::combine(prots[[3]], prots[[4]]))
unst_r3 <- filterNA(MSnbase::combine(prots[[5]], prots[[6]]))
lps_r1 <- filterNA(MSnbase::combine(prots[[7]], prots[[8]]))
lps_r2 <- filterNA(MSnbase::combine(prots[[9]], prots[[10]]))
lps_r3 <- filterNA(MSnbase::combine(prots[[11]], prots[[12]]))


tot_prots <- data.frame("Unstimulated" = c(`Replicate 1` = nrow(unst_r1), 
                                           `Replicate 2` = nrow(unst_r2), 
                                           `Replicate 3`= nrow(unst_r3)),
                        "LPS" = c(`Replicate 1` = nrow(lps_r1), 
                                                 `Replicate 2` = nrow(lps_r2), 
                                                 `Replicate 3` = nrow(lps_r3)))

library("knitr")
library("kableExtra")
kable(tot_prots, caption = "Number of quantified proteins per replicate per condition") %>%
  kable_minimal()

3.3 Protein level data processing

As shown in the above section we quantify between ~4800-5800 proteins per replicate. The below Venn diagrams show the overlap between replicates within conditions. We find 3882 proteins common in the unstimulated hyperLOPIT experiment and 4067 in the 12h-LPS stimulated hyperLOPIT experiment.

## Load R packages required for genetrating Venn diagrams
library("VennDiagram")
library("scales")
library("ggplot2")

venn.diagram(
  x = list(featureNames(unst_r1), featureNames(unst_r2), featureNames(unst_r3)),
  category.names = c("Replicate 1" , "Replicate 2 " , "Replicate 3"),
  filename = "venn_unst_replicates.png",
  main = "Unstimulated hyperLOPIT",
  col=c("#440154ff", '#21908dff', '#fde725ff'),
  fill = c(alpha("#440154ff",0.3), alpha('#21908dff',0.3), alpha('#fde725ff',0.3)),
  fontfamily = "sans",
  cat.fontfamily = "sans",
  main.fontfamily = "sans",
  cat.col = c("#440154ff", '#21908dff', '#fde725ff'), 
  output = FALSE
)

venn.diagram(
  x = list(featureNames(lps_r1), featureNames(lps_r2), featureNames(lps_r3)),
  category.names = c("Replicate 1" , "Replicate 2 " , "Replicate 3"),
  filename = "venn_lps_replicates.png",
  main = "LPS stimulated hyperLOPIT 12h-LPS",
  col=c("#440154ff", '#21908dff', '#fde725ff'),
  fill = c(alpha("#440154ff",0.3), alpha('#21908dff',0.3), alpha('#fde725ff',0.3)),
  fontfamily = "sans",
  cat.fontfamily = "sans",
  main.fontfamily = "sans",
  cat.col = c("#440154ff", '#21908dff', '#fde725ff')
)

Figure 3.1: Figure 3. Overlap of protein groups identified between replicated experiments

(Venn diagrams as per supplementary figure 1 of [1])

3.4 Dataset concatenation

We use the combine function in MSnbase to perform data fusion (concatenation) within each condition to maximise subcellular resolution and thus the reliability in protein classification [as per the pRoloc pipeline [5, 6] and in [7]]. Trotter et al. [7] have shown a significant improvement in classifier accuracy when combining replicated experiments in spatial proteomics. Specifically, we concatenate all gradient fractions across replicated experiments to give better discrimination between subcellular niches (as shown in [7]) which gives users the opportunity to resolve niches that may not have been discovered when examining replicates alone.

We note (as explained in [5, 6]) that direct comparisons of individual TMT channels in replicated experiments do not provide an adequate, goal-driven assessment of different experiments due to the nature of the gradient fraction collection, where quantitative channels do not correspond to identical selected fractions along the gradient.

Concatenate replicates for each condition,

## combine replicates for each condition
unst_full <- filterNA(MSnbase::combine(unst_r1, unst_r2))
unst_full <- filterNA(MSnbase::combine(unst_full, unst_r3)) 

lps_full <- filterNA(MSnbase::combine(lps_r1, lps_r2))
lps_full <- filterNA(MSnbase::combine(lps_full, lps_r3)) 

venn.diagram(
  x = list(featureNames(unst_full), featureNames(lps_full)),
  category.names = c("Unstimulated" , "LPS-stimulated"),
  filename = "venn_hyperLOPIT.png",
  main = "Subset of common proteins between hyperLOPIT conditions",
  col=c("#440154ff", '#21908dff'),
  fill = c(alpha("#440154ff",0.3), alpha('#21908dff',0.3)),
  fontfamily = "sans",
  cat.fontfamily = "sans",
  main.fontfamily = "sans",
  cat.col = c("#440154ff", '#21908dff'),
  margin = 0.05,
  cat.pos = c(-140, 140),
  cat.dist = c(0.05, 0.05)
)

Figure 4. Overlap of protein groups identified in both conditions

We find 3288 proteins common all 12 experiments. In the next code chunk we subset our data to keep only proteins common between all experiments.

## keep only those in common between conditions
cmn <- intersect(featureNames(unst_full), featureNames(lps_full))
lps <- lps_full[cmn, ]
unst <- unst_full[cmn, ]

lps <- addMarkers(lps, markers = "csv/markers_2019.28.01.csv")

## Markers in data: 783 out of 3288

## organelleMarkers
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    75                    73                    52 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                    69                    39                    49 
##          Mitochondria             Nucleolus               Nucleus 
##                   197                    38                    91 
##            Peroxisome       Plasma Membrane               unknown 
##                    29                    71                  2505

unst <- addMarkers(unst, markers = "csv/markers_2019.28.01.csv")

## Markers in data: 783 out of 3288

## organelleMarkers
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    75                    73                    52 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                    69                    39                    49 
##          Mitochondria             Nucleolus               Nucleus 
##                   197                    38                    91 
##            Peroxisome       Plasma Membrane               unknown 
##                    29                    71                  2505

source("r/prettymap.R")
## set organelle colours
mycol <- c("#88CCEE", "#332288", "#53CAB7", "#0170b4", "#204f20", "#990000",
           "#E69F00", "#DDCC77", "#E18493", "#AA4499", "#D55E00")
setStockcol(mycol)
setUnknownpch(16)


## Generate the t-SNE coords (Figure 4 of [1])
set.seed(399)
tsne_unst <- plot2D(unst, method = "t-SNE", plot = FALSE)
set.seed(399)
tsne_lps <- plot2D(lps, method = "t-SNE", plot = FALSE)


## plot the data as t-SNE maps
par(mfrow = c(1, 2))        
oo = c(1:3, 10, 5, 4, 6:7, 9, 8, 11)
myleg <- c(getMarkerClasses(unst), "Unknown")

prettyTSNE(tsne_unst, unst, fcol = "markers", 
          main = "Subcellular markers\nunstimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))
prettyTSNE(tsne_lps, lps, 
          main = "Subcellular markers\n12h-LPS stimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))

## add a legend
plot(NULL, xaxt='n',yaxt='n',bty='n',ylab='',xlab='', xlim=0:1, ylim=0:1)
legend("topleft", legend = myleg, col = mycol, bty = "n",
       pch = 19, cex = 1.2, pt.cex = 1.6)

3.5 Classification using a Bayesian framework

Proteins are assigned to one of the 11 subcellular marker classes using TAGM-MCMC [9]. TAGM-MCMC is a semi-supervised Bayesian generative classifier based on a T-Augmented Gaussian Mixture model that uses Bayesian computation performed using Markov-chain Monte Carlo. It is one of the many supervised machine learning methods available in the pRoloc package, but currently the only Bayesian method that is available in the package for the prediction of subcellular location for spatial proteomics data.

The advantage of using TAGM, and Bayesian methods in general, over classic supervised machine learning approaches is the ability for the algorithm to quantify the localisation uncertainty. The TAGM classifier models each annotated subcellular class using a Gaussian distribution and thus the full dataset can be modelled as a mixture of Gaussians. As described in [9] an outlier component probability is also included in this model which is mathematically described as a multivariate student’s t-distribution which gives extra information regarding the

Full details of the TAGM Bayesian methodology and mathematics are found in [9] and a detailed Bioconductor workflow for Bayesian analysis of spatial proteomoics data can be found in [10]. We follow this workflow for the analysis used here.

## Load the coda library for calculating the Gelman  diagnostics
library("coda")

## Train TAGM-MCMC
p_lps <- tagmMcmcTrain(object = lps,
                       numIter = 25000,
                       burnin = 5000,
                       thin = 10,
                       numChains = 9)

p_unst <- tagmMcmcTrain(object = unst,
                        numIter = 25000,
                        burnin = 5000,
                        thin = 10,
                        numChains = 9)

## All chains look good and all oscillate around an average of 490 outliers
out_unst <- mcmc_get_outliers(p_unst)
out_lps <- mcmc_get_outliers(p_lps)

## Calculate Gelman's Diagnostic
gelman.diag(out_unst)   
gelman.diag(out_lps)

# check all pairs as per the f1000 vignette [10]
gelman.diag(out_unst[1:2])
gelman.diag(out_unst[1:3]) # etc...

mcmc_unst_pooled <- mcmc_pool_chains(mcmc_unst)
mcmc_lps_pooled <- mcmc_pool_chains(mcmc_lps)

mcmc_unst_pooled <- tagmMcmcProcess(mcmc_unst_pooled)
mcmc_lps_pooled <- tagmMcmcProcess(mcmc_lps_pooled)

summary(mcmc_unst_pooled@summary@posteriorEstimates)
summary(mcmc_unst_pooled@summary@tagm.joint)

summary(mcmc_lps_pooled@summary@posteriorEstimates)
summary(mcmc_lps_pooled@summary@tagm.joint)

mcmc_lps <- mcmc_lps_pooled
mcmc_unst <- mcmc_unst_pooled

unst <- tagmMcmcPredict(unst, params = mcmc_unst_pooled, probJoint = TRUE)
lps <- tagmMcmcPredict(lps, params = mcmc_lps_pooled, probJoint = TRUE)

## Load the data from pRolocdata
data("thpLOPIT_unstimulated_mulvey2021")
data("thpLOPIT_lps_mulvey2021")

## We subset for proteins common in all experiments
cmn <- intersect(featureNames(thpLOPIT_unstimulated_mulvey2021),
                 featureNames(thpLOPIT_lps_mulvey2021))
unst <- thpLOPIT_unstimulated_mulvey2021[cmn, ]
lps <- thpLOPIT_lps_mulvey2021[cmn, ]


## Generate the t-SNE coords 
set.seed(399)
tsne_unst <- plot2D(unst, method = "t-SNE", plot = FALSE)
set.seed(399)
tsne_lps <- plot2D(lps, method = "t-SNE", plot = FALSE)

## plot TAGM assignments
par(mfrow = c(1, 2))
mycol <- paste0(getStockcol())
oo <- c(1:3, 10, 5, 4, 6:7, 9, 8, 11)
myleg <- c(getMarkerClasses(unst), "Unknown")

prettyTSNE(tsne_unst, unst, fcol = "tagm.mcmc.allocation", 
          main = "TAGM-MCMC allocation\nunstimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))
prettyTSNE(tsne_lps, lps, fcol = "tagm.mcmc.allocation",
          main = "TAGM-MCMC allocation\n12h-LPS stimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))

plot(NULL, xaxt='n',yaxt='n',bty='n',ylab='',xlab='', xlim=0:1, ylim=0:1)
legend("topleft", legend = myleg, col = mycol, bty = "n",
       pch = 19, cex = 1.2, pt.cex = 1.6)

   localisation probability = posterior probability * outlier probability  

fData(unst)$localisation.prob <- fData(unst)$tagm.mcmc.probability * (1 - fData(unst)$tagm.mcmc.outlier)
fData(lps)$localisation.prob <- fData(lps)$tagm.mcmc.probability * (1 - fData(lps)$tagm.mcmc.outlier)

## localisation probability = posterior probability * outlier probability
fData(unst)$localisation.prob <- fData(unst)$tagm.mcmc.probability * (1 - fData(unst)$tagm.mcmc.outlier)
fData(lps)$localisation.prob <- fData(lps)$tagm.mcmc.probability * (1 - fData(lps)$tagm.mcmc.outlier)

## Threshold = (posterior > 0.99) * 1 - (outlier < 1e-6)
t_loc <- 0.99 * (1 - 1e-6)
unst <- getPredictions(unst, 
                       fcol = "tagm.mcmc.allocation", 
                       scol = "localisation.prob", 
                       t = t_loc)

## ans
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    83                   153                   574 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                   352                    47                   261 
##          Mitochondria             Nucleolus               Nucleus 
##                   477                    39                   315 
##            Peroxisome       Plasma Membrane               unknown 
##                    34                   165                   788

lps <- getPredictions(lps, 
                      fcol = "tagm.mcmc.allocation", 
                      scol = "localisation.prob", 
                      t = t_loc)

## ans
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    91                   103                   440 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                   387                    43                   266 
##          Mitochondria             Nucleolus               Nucleus 
##                   475                    39                   385 
##            Peroxisome       Plasma Membrane               unknown 
##                    40                   227                   792

## plot TAGM assignments
par(mfrow = c(1, 2))
mycol <- paste0(getStockcol())
oo = c(1:3, 10, 5, 4, 6:7, 9, 8, 11)
myleg <- c(getMarkerClasses(unst), "Unknown")

prettyTSNE(tsne_unst, unst, fcol = "tagm.mcmc.allocation.pred", 
          main = "TAGM-MCMC final assignment\nunstimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))

## Warning in plot.window(...): "oulineCol" is not a graphical parameter

## Warning in plot.xy(xy, type, ...): "oulineCol" is not a graphical parameter

## Warning in axis(side = side, at = at, labels = labels, ...): "oulineCol" is not
## a graphical parameter

## Warning in axis(side = side, at = at, labels = labels, ...): "oulineCol" is not
## a graphical parameter

## Warning in box(...): "oulineCol" is not a graphical parameter

## Warning in title(...): "oulineCol" is not a graphical parameter

prettyTSNE(tsne_lps, lps, fcol = "tagm.mcmc.allocation.pred",
          main = "TAGM-MCMC final assignment\n12h-LPS stimulated THP1-cells",
          orgOrder = oo, mainCol = mycol, oulineCol = darken(mycol))

## Warning in plot.window(...): "oulineCol" is not a graphical parameter

## Warning in plot.xy(xy, type, ...): "oulineCol" is not a graphical parameter

## Warning in axis(side = side, at = at, labels = labels, ...): "oulineCol" is not
## a graphical parameter

## Warning in axis(side = side, at = at, labels = labels, ...): "oulineCol" is not
## a graphical parameter

## Warning in box(...): "oulineCol" is not a graphical parameter

## Warning in title(...): "oulineCol" is not a graphical parameter

plot(NULL, xaxt='n',yaxt='n',bty='n',ylab='',xlab='', xlim=0:1, ylim=0:1)
legend("topleft", legend = myleg, col = mycol, bty = "n",
       pch = 19, cex = 1.2, pt.cex = 1.6)

## Unstimulated protein profiles
par(mfrow = c(4, 3))
cl <- getMarkerClasses(unst)
for (i in seq(cl)) {
  par(mar = c(8, 4, 8, 2), las = 2)
  plotDist(unst[ fData(unst)$markers == getMarkerClasses(unst)[i], ], 
           pcol = paste0(getStockcol()[i], 90), xlab = "", ylim = c(0, .9))
  title(main = cl[i])
}

## LPS protein profiles
par(mfrow = c(4, 3))
cl <- getMarkerClasses(lps)
for (i in seq(cl)) {
  par(mar = c(8, 4, 8, 2), las = 2)
  plotDist(lps[ fData(lps)$markers == getMarkerClasses(lps)[i], ], 
           pcol = paste0(getStockcol()[i], 90), xlab = "", ylim = c(0, .9))
  title(main = cl[i])
}

loc_unst <- getMarkers(unst, "tagm.mcmc.allocation.pred")

## organelleMarkers
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    83                   153                   574 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                   352                    47                   261 
##          Mitochondria             Nucleolus               Nucleus 
##                   477                    39                   315 
##            Peroxisome       Plasma Membrane               unknown 
##                    34                   165                   788

loc_lps <- getMarkers(lps, "tagm.mcmc.allocation.pred")

## organelleMarkers
##      40S/60S Ribosome             Chromatin               Cytosol 
##                    91                   103                   440 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##                   387                    43                   266 
##          Mitochondria             Nucleolus               Nucleus 
##                   475                    39                   385 
##            Peroxisome       Plasma Membrane               unknown 
##                    40                   227                   792

kable(rbind("Unstimulated" = table(loc_unst), "LPS" = table(loc_lps)))

source("R/intersection.R")
## Add a label to the LPS dataset so when we combine it  
## with the unstimulated the information is not lost
.lps <- lps
.lps <- updateFvarLabels(.lps, "LPS")
cmb <- MSnbase::combine(unst, .lps)  # now combine all results

df_all <- compareDatasets(cmb, 
                          fcol1 = "localisation.pred",
                          fcol2 = "localisation.pred.LPS")
ggheatmap(df_all, title = "TAGM locations between conditions")

3.6 Translocations

One of the main aims of the THP-1 study in [1] was to use the hyperLOPIT technology to investigate the (spatial) changes in protein localisation between the unstimulated and 12h-LPS stimulated conditions. In this section we look at how the assigned location of proteins differ between the two conditions. We examine their assigned location from TAGM and also the difference between their joint posterior probability distributions.

Re-localisations i.e. proteins that do not localise to the same subcellular niche in both conditions, were extracted and categorised as one of four different translocation events:

1. Type 1: organelle-to-organelle - cases where proteins are assigned to different subcellular classes in each condition i.e. those that move from one discrete location to another.
2. Type 2: unknown-to-organelle - cases where proteins in the unstimulated data are labelled as “unknown” but are localised to one of the 11 organelle classes in the 12h-LPS stimulated data.
3. Type 3: organelle-to-unknown - cases where proteins are assigned a location to one of the 11 organelle classes in the unstimulated data but in the 12h-LPS data are labelled as “unknown”.
4. Type 4: unknown-to-unknown - dynamic proteins that were not assigned a subcellular niche. These are proteins which have a difference in their joint probability distribution equal to 1 (as calculated by the natural L2 distance, also known as the Euclidean norm).

The L2 distance is denoted by \[ d_{_{L2}norm}(x, y) = \sqrt {\sum_{i = 1}^{n}{(x_i - y_i)^2}} \] where x and y are the posterior probabilities for for the unstimulated and 12h-LPS stimulated respectively, for each ith class.

As already mentioned, proteins labelled as “unknown” location, describe proteins that did not meet the classifier threshold to be assigned to one of the 11 discrete organelle categories, as defined in our marker set. We not only wish to examine proteins that move between discrete locations (which we call type 1 translocations above) but we also wish to examine cases where in one condition the protein has been given a discrete location, and in another condition it has been described as “unknown”, and therefore not belonging to one of the 11 subcellular classes. It is important to note that in terms of machine learning, the term “unknown” is not a class in the training set, it is simply as term we use in our pipeline to describe proteins that we can say as not confidently assigned to one of the 11 classes.

These cases are still of great interest as although we may not know where the protein exactly resides in this snapshot of time we do know it is not classed to the same location in the other condition. Unknown cases encompass a number of scenarios for example, a protein having a mixed population and moonlighting between organelles, or it be that the protein localised to a subcellular niche that does not appear in our training data. Whichever reason what we are interested in is not only a change in localion but the size of the probability change, as this implies a change in localisation regardless of organelle residency.

As an extra measure of stringency we enforce that in cases 2-4 any protein considered as “unknown” must also have a high outlier probability (as per the below code chunk out = 1e-3) and thus is clearly not a member of any of the 11 classes.

For every protein the natural L2 distance (Euclidean norm) was calculated between the TAGM joint posterior probability distributions and translocation events were further filtered and ranked based on this distance to give a succinct subset of translocations for further analysis and data mining.

In the below code chunk we call the function getTranslocations which extracts the localisations of every protein in each condition then assigns a translocation type as described above and then calculates the L2 distance between the posterior probability distributions. This information is appended to the fData slot of our MSnSet along with all the other machine learning results.

source("R/getmovers.R")
extract_movers <- getTranslocations(unst, lps, 
                                    fcol = "tagm.mcmc.allocation.pred", 
                                    out = 1e-3)
unst <- extract_movers[[1]]
lps <- extract_movers[[2]]

getMarkers(unst, "translocations")

## organelleMarkers
## type1 type2 type3 type4 
##   112    62    49    30

We find 112 type1, 62 type 2, 49 type 3 and 30 type 4 translocation events. In the next section we can further examine and plot these events using alluvial and circos plots.

source("R/circos.R")

## generate my colour scheme
circos_cols <- c(getStockcol(), "grey")
orgs <- c(union(getMarkerClasses(unst), getMarkerClasses(lps)), "unknown")
(colscheme <- setNames(circos_cols, orgs))  # check levels consistent 

##      40S/60S Ribosome             Chromatin               Cytosol 
##             "#88CCEE"             "#332288"             "#53CAB7" 
## Endoplasmic Reticulum       Golgi Apparatus              Lysosome 
##             "#0170b4"             "#204f20"             "#990000" 
##          Mitochondria             Nucleolus               Nucleus 
##             "#E69F00"             "#DDCC77"             "#E18493" 
##            Peroxisome       Plasma Membrane               unknown 
##             "#AA4499"             "#D55E00"                "grey"

## get indices of translocating proteins
tl <- !is.na(fData(unst)$translocations)

## set circos plot parameters
par(mar = c(2, 2, 2, 2), cex = .5)
circos.clear()
circos.par(gap.degree = 4)

## plot circos
customChord(unst[tl, ], lps[tl, ], 
            cols = colscheme,  
            diffHeight  = -0.02, 
            transparency = 0.3, 
            link.sort = FALSE,
            labels = TRUE)

source("r/riverplot.R")
library(ggalluvial)
thp_alluvial <- riverplot(unst[tl, ], lps[tl, ], 
                          cols = colscheme,
                          labels = TRUE)
thp_alluvial + ggtitle("Translocations following 12h-LPS stimulation in THP-1 cells") 

df <- makedf(unst[tl, ], lps[tl, ],
            fcol1 = "localisation.pred",
            fcol2 = "localisation.pred",
            mrkCol1 = "markers",
            mrkCol2 = "markers")

## `summarise()` has grouped output by 'x'. You can override using the `.groups` argument.

names(df) <- c("Unstimulated", "LPS", "Number of translocations")
kable(df)

source("r/foi.R")
par(mfrow = c(1, 2))

## Unstimulated
prettyTSNE_overlay(tsne_unst, unst, 
                   fcol = "localisation.pred", 
                   main = "Type 1, 2, 3, or 4 translocation events", orgOrder =  oo, 
                   mainCol = paste0(lighten(mycol), 30),
                   oulineCol = paste0(lighten(mycol), 30))
highlightOnPlot(object = tsne_unst, pch = 21, cex = 1.7,
                col = mycol[2],  bg = lighten(lighten("grey")),
                foi = type1, lwd = 2,
                args = list(unst))
highlightOnPlot(object = tsne_unst, pch = 24, cex = 1.7,
                col = darken(mycol[7]), bg = lighten(lighten("#f4f4c8")), 
                foi = type2, lwd = 2, 
                args = list(unst))
highlightOnPlot(object = tsne_unst, pch = 21, cex = 1.7,
                col = darken("purple"), bg = lighten(lighten("#f2a6e3")), 
                foi = type3, lwd = 2,
                args = list(unst))
highlightOnPlot(object = tsne_unst, pch = 24, cex = 1.7,
                col = "#1c1c78", bg = lighten(lighten("#5fa8cc")), 
                foi = type4, lwd = 2,
                args = list(unst))

prettyTSNE_overlay(tsne_lps, lps, fcol = "localisation.pred", 
                   main = "", orgOrder =  oo,
                   mainCol = paste0(lighten(mycol), 30), 
                   oulineCol = paste0(lighten(mycol), 30))
## 12h-LPS 
highlightOnPlot(object = tsne_lps, pch = 21, cex = 1.7,
                col = mycol[2],  bg = lighten(lighten("grey")),
                foi = type1, lwd = 2,
                args = list(lps))
highlightOnPlot(object = tsne_lps, pch = 24, cex = 1.7,
                col = darken(mycol[7]), bg = lighten(lighten("#f4f4c8")), 
                foi = type2, lwd = 2, 
                args = list(lps))
highlightOnPlot(object = tsne_lps, pch = 21, cex = 1.7,
                col = darken("purple"), bg = lighten(lighten("#f2a6e3")), 
                foi = type3, lwd = 2,
                args = list(lps))
highlightOnPlot(object = tsne_lps, pch = 24, cex = 1.7,
                col = "#1c1c78", bg = lighten(lighten("#5fa8cc")), 
                foi = type4, lwd = 2,
                args = list(lps))

plot(NULL, xaxt='n',yaxt='n',bty='n',ylab='',xlab='', xlim=0:1, ylim=0:1)
legend("topleft", legend = c("Type 1: organelle-to-organelle", 
                             "Type 2: unknown-to-organelle", 
                             "Type 3: organelle-to-unknown", 
                             "Type 4: unknown-to-unknown"), 
       col = c(mycol[2], darken(mycol[7]), darken("purple"), "#1c1c78"), 
       pt.bg = c(lighten(lighten("grey")), lighten(lighten("#f4f4c8")), 
                 lighten(lighten("#f2a6e3")), lighten(lighten("#5fa8cc"))),
       bty = "n",
       pch = c(21, 24, 21, 24), pt.cex = 2.2, cex = 1.4)